Open kaushal-gawri9899 opened 1 month ago
If I understand your request correctly, I am working on effectively the same thing. It looks like your method is much more involved, so you might have better results with it. I'm working on cleaning up the code and once it's ready I'll submit a PR for it (I'm a full-time programmer with a day job, which means ~60 hrs/wk... so I'm finding the time when I can). Already submitted a PR to prep some changes for it. See #139.
I guess it's similar but based on the PR, I'm under the impression that you're trying to propagate the speaker representations using "input_values" in the encoder, right? I'm trying to use a different approach where i train the model to consider the speaker reference voice in the decoder (Causal LM) so I had tweaked the architecture as stated above.
Somehow I feel this creates the problem of having a third input competing with text_description. What comes first? description or speaker embedding? speaker embedding should just give the nuance of the specific voice. i would guess this could be handled with appropriate training. text_description could handle accent, question/instruction/statement, tone: mild/aggresive...
Hey, is it possible to allow voice cloning by implementing a two way process for encoding? Basically, before encoding, can we inject a speaker embedding to be used at time of encoding instead of solely depending on the style prompt? I'm looking to control the styling through a two way process where i can provide the required speaker embedding to the encoder for tone coloring/voice cloning and can do the rest of the styling through the prompt (ignoring who the speaker is)?